Advanced Kernel Methods for Medical Imaging
The goal of this project is to develop kernel-based machine learning methods for image classification that employ similarity measures comparing images in a hierarchical fashion - as humans do, but with the accuracy of a computer. These methods shall allow to solve challenging medical imaging problems, in particular they will be applied to the diagnosis of osteoarthritis (OA) and breast cancer, which are ranked among the most burdening diseases.
Looking at the visual cortex, it becomes obvious that the human visual system uses "deep" structure consisting of multiple levels of processing operating on more and more abstract representations of the visual scene. This has been successfully copied in computer vision systems, In contrast, kernel-based learning algorithms such as support vector machine (SVM) classifiers mark the state-of-the art in pattern recognition. They employ (Mercer) kernel functions to implicitly define a metric feature space for processing the input data, that is, the kernel defines the similarity between observations, in our case between medical images. Kernel methods are well understood theoretically and give excellent results in practice. However, they are usually considered to be "shallow" learning methods in the sense that they realize only a single layer of non-linear processing. This project will combine hierarchical image processing with the efficiency, theoretical beauty, and accuracy gain of SVMs for advancing the performance of medical imaging systems. This is made possible by marrying the applicants expertise in kernel-based machine learning with the widely recognized knowledge in medical image analysis at his new affiliation The Image Group at the Department of Computer Science, University of Copenhagen (DIKU).
The project is funded by the European Commission by a Marie Curie Career Integration Grant (PCIG10-GA-2011-303655).
Pengfei Diao. Hierarchical Kernel Methods for Medical Image Analysis. Master's thesis, Department of Computer Science, University of Copenhagen, November 2012
Peer-reviewed contributions to journals and conferences
Michiel Kallenberg, Kersten Petersen, Mads Nielsen, Andrew Y. Ng, Pengfei Diao, Christian Igel, Celine M. Vachon, Katharina Holland, Rikke Rass Winkel, Nico Karssemeijer, and Martin Lillholm. Unsupervised deep learning applied to breast density segmentation and mammographic risk scoring. IEEE Transactions on Medical Imaging, 35(5), pp. 1322-1331, 2016
Søren Frejstrup Maibing and Christian Igel. Computational Complexity of Linear Large Margin Classification With Ramp Loss. JMLR W&CP38 (AISTATS), pp. 259-267, 2015
Asja Fischer and Christian Igel. Training Restricted Boltzmann Machines: An Introduction. Pattern Recognition47, pp. 25-39, 2014
Søren Dahlgaard, Christian Igel, and Mikkel Thorup. Nearest Neighbor Classification Using Bottom-k Sketches. IEEE International Conference on Big Data 2013, pp. 28-34, IEEE Press, 2013
Adhish Prasoon, Kersten Petersen, Christian Igel, Francois Lauze, Erik Dam, and Mads Nielsen. Voxel Classification Based on Triplanar Convolutional Neural Networks Applied to Cartilage Segmentation in Knee MRI. In: Medical Image Computing and Computer Assisted Intervention (MICCAI 2013), LNCS, 8150, pp 246-253, Springer-Verlag, 2013
Adhish Prasoon, Christian Igel, Marco Loog, Francois Lauze
, Erik Dam, and Mads Nielsen
. Femoral Cartilage Segmentation in Knee MRI Scans Using Two Stage Voxel Classification. In: 35th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC)
, pp 5469-5472, IEEE Press
Christian Igel. A Note on Generalization Loss When Evolving Adaptive Pattern Recognition Systems. IEEE Transactions on Evolutionary Computation17(3), pp. 345-352, 2013
Ürün Dogan, Tobias Glasmachers, and Christian Igel. A Note on Extending Generalization Bounds for Binary Large-margin Classifiers to Multiple Classes. In P. A. Flach, T. De Bie, and N. Cristianini, eds.: European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML-PKDD 2012), LNCS 7523, pp.122-129, Springer-Verlag, 2012